The Risk of Generator-Induced Covariate Shift: Neural Network Surrogates of Elastic-Plastic Behavior

Neural network surrogates are rapidly transforming engineering simulations by replacing computationally expensive constitutive models. While surrogate modeling literature focuses extensively on architectural modifications and physical constraints, the broader machine learning community is increasingly shifting attention toward training data quality and generation strategies. Aligning with this paradigm, this study exposes a field characteristic OOD vulnerability in constitutive surrogate modeling. Training data is necessarily produced by a stochastic loading history generator, yet it is routinely overlooked that any given generation algorithm and its configuration sample exclusively from a single, fixed trajectory distribution. Neural networks are particularly effective at learning such distributional structure, therefore they risk adapting to the generator's statistical fingerprint rather than the underlying invariant constitutive law. More critically, when validation data is drawn from the same distribution, performance appears satisfactory, creating a false sense of security. We systematically analyze this phenomenon using six 1D elastoplastic constitutive models, a hyperparameter sweep of 184 neural network configurations (AR-MLPs and LSTMs ranging from 15 to 1100 parameters), and four stochastic loading history generators (including two newly proposed multisine algorithms). Our analysis yields two primary insights: (i) on a representative case, we demonstrate that out-of-distribution degradation scales with increasing model capacity (which in our investigated setup reached generalization ratios of up to approximately 10, while in-distribution performance remains near unity), and (ii) cross-generator evaluation across all material models confirms that architectural optimization on a single dataset does not guarantee cross-generator robustness, as retraining consistently triggers generator-specific adaptation with generalization ratios surging up to 7-fold. To expose these hidden failures, we introduce a rigorous, multi-scenario evaluation protocol incorporating critical loading benchmarks. This work demonstrates that training on a single loading history generator is a fundamental bottleneck, and advocates for a data-centric perspective and algorithmic data diversity as a necessary paradigm shift for training elastoplastic neural network surrogates capable of stable integration into non-linear finite element solvers.